FRE-GAN : Full-resolution efficient convolutional generative adversarial network for retinal vessel segmentation

Accurate segmentation of retinal blood vessels is not only related to clinical diagnosis and surgical planning of diseases, but also closely related to disease research. However, due to the varying thickness and low contrast of blood vessels in medical images, and constrained by the complex spatial topology, deep learning-based blood vessel segmentation methods are still quite challenging while achieving excellent performance. In order to achieve accurate retinal vessel segmentation, we propose FRE-GAN, a full-resolution efficient convolutional generative adversarial network, to achieve more efficient feature extraction and segmentation. Specifically, we design a full-resolution parallel convolutional interactive generator and the lightweight dual-domain discriminator to enrich the contextual feature information through multi-scale feature learning of image information and feature fusion of inter-neighborhood information. In addition, considering the topology of blood vessels, we reconstruct the topological continuity loss function to improve the generation quality and topological continuity of blood vessel structures. We evaluate the proposed method on the retinal vascular image datasets (DRIVE, CHASE_DB1, and STARE). The results show that with the small number of parameters, compared with the classical and state-of-the-art methods, FRE-GAN is demonstrated to achieve higher accuracy, continuity, and comprehensive performance in retinal vascular segmentation, both from the qualitative and quantitative perspectives.